WO2014150421A1 - Estimation de mouvements en cascade de caméra, détection de roulement d'obturateur, et détection de tremblement de caméra pour améliorer la stabilisation d'une vidéo - Google Patents

Estimation de mouvements en cascade de caméra, détection de roulement d'obturateur, et détection de tremblement de caméra pour améliorer la stabilisation d'une vidéo Download PDF

Info

Publication number
WO2014150421A1
WO2014150421A1 PCT/US2014/023211 US2014023211W WO2014150421A1 WO 2014150421 A1 WO2014150421 A1 WO 2014150421A1 US 2014023211 W US2014023211 W US 2014023211W WO 2014150421 A1 WO2014150421 A1 WO 2014150421A1
Authority
WO
WIPO (PCT)
Prior art keywords
homographic
video
motion
model
frame
Prior art date
Application number
PCT/US2014/023211
Other languages
English (en)
Inventor
Matthias Grundmann
Vivek Kwatra
Irfan ESSA
Original Assignee
Google Inc.
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Google Inc. filed Critical Google Inc.
Priority to EP20209572.5A priority Critical patent/EP3800878B1/fr
Priority to EP14768016.9A priority patent/EP2974274B1/fr
Priority to KR1020157029569A priority patent/KR102185963B1/ko
Priority to CN201480016067.2A priority patent/CN105052129B/zh
Priority to KR1020207034211A priority patent/KR102225410B1/ko
Publication of WO2014150421A1 publication Critical patent/WO2014150421A1/fr

Links

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N23/00—Cameras or camera modules comprising electronic image sensors; Control thereof
    • H04N23/60—Control of cameras or camera modules
    • H04N23/68—Control of cameras or camera modules for stable pick-up of the scene, e.g. compensating for camera body vibrations
    • H04N23/682—Vibration or motion blur correction
    • H04N23/683—Vibration or motion blur correction performed by a processor, e.g. controlling the readout of an image memory
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/20—Analysis of motion
    • G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • G06V20/46—Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N23/00—Cameras or camera modules comprising electronic image sensors; Control thereof
    • H04N23/60—Control of cameras or camera modules
    • H04N23/68—Control of cameras or camera modules for stable pick-up of the scene, e.g. compensating for camera body vibrations
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N23/00—Cameras or camera modules comprising electronic image sensors; Control thereof
    • H04N23/60—Control of cameras or camera modules
    • H04N23/68—Control of cameras or camera modules for stable pick-up of the scene, e.g. compensating for camera body vibrations
    • H04N23/681—Motion detection
    • H04N23/6811—Motion detection based on the image signal
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N23/00—Cameras or camera modules comprising electronic image sensors; Control thereof
    • H04N23/60—Control of cameras or camera modules
    • H04N23/68—Control of cameras or camera modules for stable pick-up of the scene, e.g. compensating for camera body vibrations
    • H04N23/689—Motion occurring during a rolling shutter mode
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10016—Video; Image sequence
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20021—Dividing image into blocks, subimages or windows
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30244—Camera pose

Definitions

  • This disclosure generally relates to manipulating video content and more specifically to stabilizing camera motion in video content.
  • rolling shutter is corrected at least in part by
  • a homographic model is applied to the inter frame motion to determine a number of tracked features that are inliers matching the homographic model.
  • a number of homographic mixture models are applied to the tracked features to determine, for each of the homographic mixture models, a number of tracked features that are inliers matching each homographic mixture model, the homographic mixture models having different rigidities.
  • a least rigid homographic mixture model is determined where the number of homographic mixture inliers exceeds the number of homographic inliers by a threshold specific to that homographic mixture model.
  • a stabilized video is generated by applying the least rigid homographic mixture model to the adjacent frames of the video.
  • a video is classified as being likely to benefit from
  • a spectrogram is generated for each of the DOFs, each spectrogram based on the values of the DOFs over a time window comprising a plurality of adjacent frames of the video.
  • a plurality of shake features are generated based on the spectrograms. The video is classified based on the shake features. The video is then stabilized based on the classification.
  • FIG. 1 is a high-level block diagram of a computing environment including a video stabilization system, according to one embodiment.
  • FIG. 2 is a high-level block diagram illustrating an example of a computer for use as a video stabilization system , video server, and/or client.
  • FIG. 3 is a high-level block diagram illustrating modules within the video stabilization system, according to one embodiment.
  • FIG. 4 is a flowchart illustrating a process for determining a camera path of a video, according to one embodiment.
  • FIG. 5 is a flowchart illustrating a process for detecting and correcting rolling shutter in a video, according to one embodiment.
  • FIG. 6 illustrates motion of example tracked features and their motion within a frame, according to one embodiment.
  • FIG. 7 illustrates a number of motion models each having a different number of degrees of freedom, according to one embodiment.
  • FIG. 8 is a flowchart illustrating a process for detecting camera shake in a video, according to one embodiment.
  • FIGs. 9A and 9B illustrate a number of spectrograms for a number of time windows and for the different degrees of freedom of the similarity model, where FIG. 9A illustrates the spectrograms of a first, short-length video 12 windows in length, and FIG. 9B illustrates the spectrograms of a second, longer-length video 40 windows in length, according to one embodiment.
  • FIG. 1 is a high-level block diagram of a computing environment including a video stabilization system, according to one embodiment.
  • FIG. 1 illustrates a video server 110, a video stabilization system 112 (the "stabilization system") and a client 114 connected by a network 116. Only one client 114 is shown in FIG. 1 in order to simplify and clarify the description.
  • Embodiments of the computing environment 100 can have thousands or millions of clients 114, as well as multiple video server 110 and stabilization systems 112.
  • the video server 110 serves video content (referred to herein as "videos") to clients 114 via the network 116.
  • the video server 110 is located at a website provided by YOUTUBETM, although the video server can also be provided by another entity.
  • the video server 110 includes a database storing multiple videos and a web server for interacting with clients 114.
  • the video server 110 receives requests from users of clients 114 for the videos in the database and serves the videos in response.
  • the video server 110 can receive, store, process (e.g., stabilize) and serve videos posted by users of the clients 114 and by other entities.
  • the client 114 is a computer or other electronic device used by one or more users to perform activities including uploading videos, initiating the stabilization of videos using the stabilization system 112, and viewing videos and other content received from the video server 110.
  • the client 114 can be a personal computer executing a web browser 118 that allows the user to browse and search for videos available at the video server web site.
  • the client 114 is a network-capable device other than a computer, such as a personal digital assistant (PDA), a mobile telephone, a pager, a television "set-top box,” etc.
  • PDA personal digital assistant
  • the network 116 enables communications among the entities connected to it.
  • the network 116 is the Internet and uses standard communications technologies and/or protocols.
  • the network 116 can include links using technologies such as Ethernet, 802.11, worldwide interoperability for microwave access (WiMAX), 3G, digital subscriber line (DSL), asynchronous transfer mode (ATM), InfiniBand, PCI Express Advanced Switching, etc.
  • the networking protocols used on the network 116 can include multiprotocol label switching (MPLS), the transmission control protocol/Internet protocol (TCP/IP), the User Datagram Protocol (UDP), the hypertext transport protocol (HTTP), the simple mail transfer protocol (SMTP), the file transfer protocol (FTP), etc.
  • MPLS multiprotocol label switching
  • TCP/IP transmission control protocol/Internet protocol
  • UDP User Datagram Protocol
  • HTTP hypertext transport protocol
  • SMTP simple mail transfer protocol
  • FTP file transfer protocol
  • the data exchanged over the network 116 can be represented using technologies and/or formats including the hypertext markup language (HTML), the extensible markup language (XML), etc.
  • HTML hypertext markup language
  • XML extensible markup language
  • all or some of links can be encrypted using conventional encryption technologies such as the secure sockets layer (SSL), transport layer security (TLS), virtual private networks (VPNs), Internet Protocol security (IPsec), etc.
  • SSL secure sockets layer
  • TLS transport layer security
  • VPNs virtual private networks
  • IPsec Internet Protocol security
  • the entities use custom and/or dedicated data communications technologies instead of, or in addition to, the ones described above.
  • the stabilization system 112 is configured to receive an input video and to stabilize it by altering the pixel content of the frames of the video.
  • the stabilization system 112 outputs a stabilized video.
  • the stabilization system 112 determines a camera path describing the two dimensional (2D) motion of the camera originally used to record the video.
  • the stabilization system 112 may also output this camera path separately from using it merely to stabilize the video.
  • the camera path is used to negate, to the extent possible, the motion of pixels in the frames of the video due to the motion of the camera.
  • the output stabilized video is a copy of the original video where the positions of the pixels of each frame are adjusted to counteract the motion between frames according to the determined camera path.
  • FIG. 2 is a high-level block diagram illustrating an example of a computer 200 for use as a video server 110, stabilization system 112, and/or client 114. Illustrated are at least one processor 202 coupled to a chipset 204.
  • the chipset 204 includes a memory controller hub 220 and an input/output (I/O) controller hub 222.
  • a memory 206 and a graphics adapter 212 are coupled to the memory controller hub 220, and a display device 218 is coupled to the graphics adapter 212.
  • a storage device 208, keyboard 210, pointing device 214, and network adapter 216 are coupled to the I/O controller hub 222.
  • Other embodiments of the computer 200 have different architectures.
  • the memory 206 is directly coupled to the processor 202 in some embodiments.
  • the storage device 208 is a non-transitory computer-readable storage medium such as a hard drive, compact disk read-only memory (CD-ROM), DVD, or a solid-state memory device.
  • the memory 206 holds instructions and data used by the processor 202.
  • the pointing device 214 is a mouse, track ball, or other type of pointing device, and is used in combination with the keyboard 210 to input data into the computer system 200.
  • the graphics adapter 212 displays images and other information on the display device 218.
  • the network adapter 216 couples the computer system 200 to the network 116. Some embodiments of the computer 200 have different and/or other components than those shown in FIG. 2.
  • the computer 200 is adapted to execute computer program modules for providing functionality described herein.
  • module refers to computer program instructions and other logic used to provide the specified functionality.
  • a module can be implemented in hardware, firmware, and/or software.
  • program modules formed of executable computer program instructions are stored on the storage device 208, loaded into the memory 206, and executed by the processor 202.
  • the types of computers 200 used by the entities of FIG. 1 can vary depending upon the embodiment and the processing power used by the entity.
  • a client 114 that is a mobile telephone typically has limited processing power, a small display 218, and might lack a pointing device 214.
  • the stabilization system 112 may comprise multiple servers working together to provide the functionality described herein. As will be apparent from the description below, the operations of the stabilization system 112 to stabilize a video are sufficiently complex as to require their implementation by a computer, and thus cannot be performed entirely mentally by the human mind.
  • FIG. 3 is a high-level block diagram illustrating modules within the video stabilization system 112, according to one embodiment. As introduced above, the
  • stabilization system 112 is configured to receive an input video 302, to stabilize the video, and output the stabilized video 304 and/or the camera path 306.
  • the stabilization system 112 includes a motion estimation system module 310, a camera path analysis module 320, a camera path stabilization module 330, a stabilized video generation module 340, a data storage module 350, and a shake detection module 360.
  • Some embodiments of the stabilization system 112 have different and/or additional modules than the ones described here.
  • the functions can be distributed among the modules in a different manner than is described here. Certain modules and functions can be incorporated into other modules of the stabilization system 112 and/or other entities on the network 116, including the video server 110 and client 114.
  • the data storage module 350 stores data used by the various modules of the stabilization system 112.
  • the stored data include, for example, frames and/or other portions of videos being operated upon, tracked features, estimated motion models, properties and thresholds related to the stabilization process, camera paths, and other intermediate items of data created during the stabilization process. This list is intended to be exemplary and not exhaustive.
  • the motion estimation module 310 analyzes the frames of the input video 302 to characterize the original 2D camera motion of the camera used to capture the input video 302, and to provide that characterized output as a camera path, and is one means for performing this function. For a pair of adjacent frames I t and I t+ i representing times t and t+1 in the video, respectively, the motion estimation module 310 characterizes the camera path based on the movement of a set of tracked features T t , T t+ i from their initial locations in frame I t to their final locations in frame I t+ i.
  • the set of tracked features Tis generated from the underlying pixels of each frame, and their movement between adjacent frames is represented as a set of inter- frame motions M t .
  • the motion estimation module 310 uses the inter- frame motions M t to estimate a set of motion models F t for frame I t , where the application of the estimated motion models F t to the pixels of frame I t describes the motion of the pixels between frame I t and I t+ i.
  • the tracking module 312 is configured to generate a set of tracked features T t for each frame I t of the input video 312, and is one means for performing this function.
  • the tracked features act as markers for objects appearing in a video frame.
  • the tracking module 312 tracks the motion of individual tracked features between frame pairs to track how objects in the video move between frames.
  • the motion M t of the tracked features between a pair of adjacent frames can be analyzed to separate object motion within the frame from motion of the capturing camera.
  • the tracking module 312 generates the tracked features T t for a frame by applying a corner measure to the pixels of the frame (e.g., a Harris corner measure).
  • the corner measure generates a tracked feature at each pixel in the frame where a "corner" appears, that is, where the vertical and horizontal lines of significant gradient in pixel color meet. More specifically, the tracked features are located at pixels where the minimum eigenvalue of the auto-correlation matrix of the gradient of the frame is above a threshold after non-maxima suppression.
  • the tracked features may be stored as a set of two-dimensional (2D) points, each tracked feature having an and y axis coordinate within the Cartesian coordinate system of the frame of the video.
  • 2D two-dimensional
  • the tracking module 312 may divide the frame into multiple layers of grids having different sizes (e.g., 4x4 or 16 grids total, 8x8 grids, and 16x16 grids).
  • the gradient threshold for what may be considered a tracked feature may set on a per-grid basis to normalize the number of tracked features generated per cell of the grid. This helps balance the number of tracked features arising out of each portion of the frame, so that the tracked features are not overly representative of some cells over others. That way, cells with large amounts of color change over a relatively short distance will not necessarily have more tracked features than cells that are more uniform in color.
  • An absolute minimum threshold may be enforced to address homogeneous regions of a frame.
  • the absolute minimum threshold may conclude that particular regions may lack tracked features. Tracked features that are in close proximity to other tracked features (e.g., within 5 pixels) may be aggregated or filtered in order to ensure that the tracked features are spread out within a cell, and within the frame as a whole.
  • FIG. 6 illustrates motion of example tracked features and their motion M t within a frame, according to one embodiment.
  • the tracked features will exhibit motions that are inconsistent with the motion other nearby tracked features.
  • the tracked features are analyzed to identify and filter out these outlier tracked features.
  • Inconsistent motion can include, for example, a tracked feature T t moving M t in
  • the threshold for what represents a different direction and what represents nearby is determined several times at several different levels.
  • Several levels of grids e.g., 4x4 or 16 grids total, 8x8 grids, and 16x16 grids) are used as introduced above, where each level of grid has a different threshold for what constitutes a nearby tracked feature and what constitutes a substantially different direction.
  • tracked features within a cell of a grid are considered nearby.
  • the direction of motion of the tracked features of a cell is determined based on an aggregate (e.g., an average) of the directions of motions of each of the tracked features of that cell.
  • the threshold tolerance for a substantially different direction may be set very high (e.g., requiring a high amount of uniformity) between tracked features for larger grids (e.g., 16x16), and may be set comparatively lower (e.g., requiring less uniformity) between tracked features for smaller grids (e.g., 4x4).
  • Tracked features not meeting the directionality threshold at one or more levels are thrown out.
  • tracked features are filtered using a random sample consensus (RANSAC) algorithm.
  • RANSAC random sample consensus
  • all but one tracked feature in an example grid may exhibit leftward translation, whereas the remaining tracked feature exhibits rightward translation.
  • the rightward moving tracked feature may be filtered and not considered in further processing.
  • the cascaded motion module 314 is configured to use the set of inter- frame motions M t of the set of tracked features T t between pairs of adjacent frames to determine the original camera path, and is one means for performing this function. To do this, the cascaded motion module 314 fits the set of inter- frame motions M t to a set of linear motion models F t . Each of the motion models represents a different type of motion having a different number of degrees of freedom (DOF).
  • DOF degrees of freedom
  • the output camera path of the cascaded motion module 314 is, for each pair of frames, the estimated motion models that are determined to be valid representations of the inter- frame motions.
  • the set of motion models determined valid for a pair of adjacent frames I t and I t+1 are assigned to the second frame, I t +1 , in the pair.
  • a set of valid motion models is determined assigned for each frame in the video except the first, as there is no motion to be analyzed in the first frame of the video.
  • an identity motion model is used for initialization.
  • applying the valid motion models of frame I t is at least part of the process used to generate stabilized video frame J t+ i, which is frame I t+ i stabilized for the original camera motion.
  • the set of valid motion models may be assigned to the first frame in the pair instead.
  • FIG. 7 illustrates a number of motion models each having a different number of degrees of freedom, according to one embodiment.
  • the first motion model 0 is a translation model having two degrees of freedom for detecting motion along the x and y axes of the frame.
  • the second motion model FP is a similarity model with four degrees of freedom for detecting rotations and uniform scaling (e.g., size of the frame), as well as for detecting translation.
  • the third motion model ) is a is a homographic model having eight degrees of freedom for detecting perspective effects, skew, non-uniform scales, as well as for detecting similarities and translations.
  • the homographic mixture model detects rolling shutter distortions (e.g., wobble), in addition to detecting homographies, similarities, and translations.
  • rolling shutter distortions e.g., wobble
  • each motion model includes new degrees of freedom representing a new type of camera and also includes DOFs for the motions represented by the lower DOF motion models. Exemplary motion models are further described below.
  • the motion models are each configurable with their own parameters, where each parameter represents one of the DOF of the motion model.
  • the cascaded motion module 314 estimates the parameters of the motion models to determine the configuration of each motion model that best fits the inter-frame motions M t .
  • the estimated motion models can be evaluated to determine whether or not they are the "correct" models to be applied. That is, they are evaluated for validity to determine whether or not they represent the motion M t of tracked features between frames. This is further described below in the next section.
  • the parameters of a given motion model are determined so as to minimize:
  • each i represents an inter-frame motion between two corresponding tracked features of a frame pair
  • Inlier tracked features fitting the model have values much greater than 1 , and outliers have small weights having values close to or less than 1.
  • the parameters each motion model F t are estimated by minimizing the sum of Eq. 2.
  • the parameters (DOF) of each motion model are as follows.
  • the translation motion model F ⁇ is estimated as the weighted average translation of the tracked features with weights Wj such that:
  • f and f represent the magnitude of translation of the camera along the x and y axes, respectively.
  • the magnitude of translation may be expressed in pixels or as a percent of the frame width/height before cropping.
  • the values of f and f can also be said to represent the values of the DOFs of the translation model for that frame.
  • the similarity motion model FP is estimated such that:
  • a is a frame-constant scale parameter
  • b represents rotation
  • t are translations in x and y.
  • the homographic model is estimated using a 3x3 matrix, where up-to-scale ambiguity is resolved by normalizing the matrix elements such that equals 1.
  • the matrix elements of the homographic model are estimated using a weighted version of the non- homogeneous direct linear transformation (DLT) algorithm solved via QR decomposition.
  • DLT direct linear transformation
  • a and d are frame-constant scale parameters
  • t are translations in x and y
  • c and b are rotation and skew, respectively.
  • the homographic mixture model FP is estimated using a mixture of a number (e.g., 10) of different homographic models, as well as a regularizer that may vary in value between different implementations.
  • the homographic mixture model applies a different homographic model to each portion of the frame. More specifically, a block is a set of consecutive scan lines in a frame, where the total number of scan lines in the frame is partitioned into 10 blocks of scan lines. Thus, a different homographic model is applied to each block.
  • the regularizer affects the rigidity of the homographic mixture model. For example, a regularizer value of a sufficiently high value (e.g., 1) causes the homographic mixture model to be rigid, making it identical to the homographic model FP . A smaller regularizer value (e.g., between 0 and 1) increases the contribution of the other
  • the homographic mixture model FP is represented by
  • w T (wi, W 2 P is the frame-constant perspective part, a and d are frame-constant scale parameters, tk are block- varying translations in x and y, and and bk are block- varying rotation and skew.
  • the tracked features T t contributing to the inter- frame motions M t are filtered prior to estimation of the homographic FP and homographic mixture model FP parameters.
  • the parameters for the similarity model FP are estimated first.
  • a set of one or more tracked features not matching the estimated similarity model are determined.
  • These non-matching tracked features are filtered from use in estimating the homographic and homographic mixture model parameters, for at least the first iteration of Eq. 2. This may be accomplished, for example, by setting their weights to zero. This helps insulate the parameters of the homographic and homographic mixture models against significant foreground motions (e.g., motions very close to the camera).
  • the weights of the tracked features may, in an alternative embodiment, be biased to give greater weight to tracked features near the edge of frame, and to give less weight to tracked features near the center of the frame. This may be accomplished, for example, using an inverted Gaussian function along the x and y coordinate axes of the frame. This is based on a prediction that faces and other objects close to the camera frame tend to be centered with respect to the frame.
  • the inter-frame motions M t of tracked features T t between any given pair of frames may look like any, some, all, or none of the estimated motion models. For example, if the scene is strictly non-planar (e.g. due to different depth layers or significant foreground motions) the translation motion model will be insufficient in describing the motion (with the translation model generating the least number of stabilization artifacts relative to the other motion models). Application of the correct (or valid) set of motion models to the inter-frame motion will stabilize those frames and remove at least some destabilizations, resulting in residual shake. Application of incorrect models introduces distortions into both the camera path and the stabilized video that were not originally present.
  • the result of its application will be a reduction of shake. If the translation model is invalid, the result of its application will be additional shake distortion. If the similarity model is valid, the result of its application will be the introduction of high frequency rigid wobble residual shake (mostly perspective in nature). If the similarity model is invalid, the result of its application will be additional shake distortion. If the homographic model is valid, the result of its application will be close to none residual shake if there is no rolling shutter present, and wobble residual shake if there is rolling shutter present. If the homographic model is invalid, the result of its application will be perspective warping errors. If the homographic mixture model is valid, the result of its application will be close to none residual shake. If the homographic mixture model is invalid, the result of its application will be non-rigid wave-like warp distortions.
  • the motion models are fit to the set of tracked features T t , T t+1 and inter- frame motions M t to determine which motion models F t validly match the inter- frame motion.
  • a motion model is considered to be valid with respect to an inter- frame motion if the type of motion represented by the motion model matches the exhibited inter- frame motion with respect to one or more properties. These properties represent the degree of fit between the motion model and the inter-frame motions. The properties differ among the motion models.
  • Table 1 illustrates an example set of properties for validity evaluation, according to one embodiment.
  • Table 1 includes the motion models, the properties relevant to each motion model, and a threshold. Some properties are simply the parameters of the motion models estimated for the inter-frame motions. Others properties are derivable from the fit of the estimated motion model to the tracked features T t , T t+ i and inter- frame motions M t .
  • Tracked features matching the model may be referred to as inliers, and tracked features not matching the model may be referred to as outliers.
  • a tracked feature is an inlier if it fits the estimated motion model to within a threshold tolerance. For example, if the motion model predicted the motion M t ,i of a tracked feature T t between to within 1.5 pixels of accuracy, than the tracked feature may be considered an inlier.
  • the motion model is invalid.
  • other properties, thresholds, and requirements may be defined for determining whether or not a motion model is valid.
  • the number of tracked features is the total number of tracked feature inliers.
  • the translation magnitude is the amount of inter-frame motion estimated by the translation model. This may be determined, for example, from a translation magnitude parameter of the motion model. Standard deviation of translation may be determined based on the individual translations of the tracked features between frames. Acceleration may be determined based on the average pixel shift of the tracked features between a pair of frames relative the median of the average pixel shift from one or more previous frame pairs (e.g., 5 previous frame pairs).
  • the number of tracked features is the same as for the translation model.
  • the feature coverage as a percentage of frame area is determined by placing a box having a fixed size around each feature and by taking the union of all the boxes. The area within the union of the boxes is compared against the total frame area to determine the feature coverage.
  • the scale change and rotation properties may be determined based scale change and rotation parameters, respectively, of the similarity model.
  • the homographic properties may also include a perspective property that may be determined based on a change in perspective parameter from the homographic model.
  • the threshold for the perspective property is based on a per-normalization, is unit-less, and may, for example, be 4x10 ⁇ 4 in value.
  • the grid coverage property represents a calculation of the amount of the frame that is covered by inlier tracked features.
  • the grid coverage property is determined by overlaying a grid (e.g., 10x10) over the tracked features of the frame pair. For each cell (or bin) of the grid, a score is determined whether the bin is an inlier or outlier bin.
  • the bin score is based on whether the tracked features in the bin are inliers or outliers with respect to the homographic model, and based on the weights of the tracked features in the bin, specifically based on the median b j of the feature weights of the tracked features in the bin. In one embodiment, the score of a bin j is determined based on
  • Grid coverage G t is an average over all bin scores, such that
  • G t is too low, the grid coverage property is too low (e.g., below 30% of the bins) and thus the homographic model may be considered invalid.
  • each block of the mixture is assigned its own block coverage score. Specifically, for 1x10 grid is overlaid on the tracked features, each bin (10 bins total) corresponds to one of the blocks. Each block thus covers a number of scan lines in the frame. A score is determined for each bin/block based on the weights of the tracked features and whether or not they are inliers. A block is considered an outlier block if its coverage is below a threshold, for example 40%.
  • the adjacent outlier blocks properties indicates the number of adjacent blocks that are outliers. If too many are outliers, the property is invalid.
  • the empty block property indicates the number of blocks having few (e.g., below a threshold) or no tracked features. If too many blocks have too few tracked features, insufficient data is available to fully validate the homographic mixture, and consequently the homographic mixture model is considered invalid.
  • the motion models are estimated, and evaluated for validity with respect to the inter- frame motion in a sequenced order, starting with the translation model and increasing in number of DOF from there. If the translation model is determined to be valid, the similarity model is considered. If the similarity model is determined to be valid, the homographic model is considered, and so on. At any point, if a model is determined to be invalid, the process is stopped and the previous model/s that was/were considered valid are used as part of the camera path for that frame. If no motion model is valid, an identity motion model used which assumes the camera path did not move (e.g., no stabilization is performed). This streamlining is efficient because often if a lower DOF motion model is invalid, it is likely that the higher DOF motion models will also be invalid.
  • the camera path analysis module 320 receives the tracked features and the valid estimated motion models from the motion estimation module 310. Generally, the camera path analysis module 320 uses these inputs to address stabilization issues that occur over a longer time span than can be detected at the inter- frame time span, for example stabilization issues that occur over hundreds of milliseconds to seconds of video. The camera path analysis module 320 performs corrections by changing the estimated motion models that are considered valid on a frame by frame basis, and by flagging frames that exhibit particular characteristics. In one embodiment, the camera path analysis module 320 includes an invalidity propagation module 322, a rolling shutter correction module 324, and an overlay and blur correction module 326.
  • the invalidity propagation module 322 is configured to smooth out the camera path over longer stretches of frames for temporal stability, and is one means for performing this function. This is based on the assumption that instabilities generally occur over multiple pairs of frames rather than in between two frames. For example, if the highest DOF valid motion model at t-1 is the homographic mixture model, at t it is the similarity model, and at t+1 it is the homographic mixture model, it is unlikely that the cause of the invalidity of the higher DOF models at t occurred only within the two frame time span between the frame time t and the frame at time t+1.
  • the number of DOF of the highest DOF valid model at a given frame pair is propagated to a number of nearby frame pairs.
  • the highest DOF valid model at time t may be the similarity model.
  • the invalidity propagation module 322 compares the number of DOF of the highest DOF valid model at that preceding or subsequent time with the number of DOF of the highest DOF valid model at time t.
  • the highest valid DOF model at the previous or subsequent time is downgraded (in terms of DOF) to match the number of DOF at time t.
  • the highest DOF valid model at times t-l and t+1 would be downgraded from the homographic mixture model to the similarity model.
  • the output of the invalidity propagation module 322 is a set valid estimated motion models that is different from the set of valid motion models received from the motion estimation module 310.
  • the rolling shutter correction module 324 is configured to analyze tracking features T t , T t+1 and inter- frame motions M t to detect and correct rolling shutter distortions, and is one means for performing this function.
  • the rolling shutter correction module 324 does not require any information from the original capturing camera regarding how the video was captured, or how the camera moved during capture.
  • Rolling shutter occurs when not all parts of a frame are recorded at the same time by the camera capturing the video. While this can be a deliberately generated effect in single image capture use cases, it is generally undesirable in videos.
  • Rolling shutter can result in several different effects, including wobble, skew, smear, and partial exposure. Generally, rolling shutter effects occur as a result of an object moving quickly within the frame during frame capture, such that the object appears to wobbles, appears skewed, etc.
  • the rolling shutter correction module 324 is configured to apply the homographic model ) estimated for that frame pair to the tracked features of that frame pair.
  • a number of homographic mixture inliers are also determined in the same manner, except the homographic mixture model F t 3 - 1 is used in place of the homographic model
  • Each score is based on the number of tracked features in the bin that are inliers or outliers with respect to either the homographic model or the homographic mixture.
  • the scores are further weighted based on the weights of the tracked features in the bin, specifically based on the median b j of the feature weights of the tracked features in the bin.
  • the score of a bin j for either case is determined based on
  • Two grid coverages are determined, a homographic grid coverage G T and a homographic mixture grid coverage each based on their respective bin scores.
  • the homographic mixture model models rolling shutter better than the homographic model. Consequently, generally the homographic mixture has a higher grid coverage GP than the homographic model's grid coverage G ) when a rolling shutter effect is present.
  • the rolling shutter correction module uses a rolling shutter boost estimate rse t to detect a rolling s
  • rolling shutter effects occur over multiple frames (e.g., on the order of hundreds of milliseconds to seconds).
  • the rolling shutter correction module 324 determines the boost rse t for multiple times/frames t (e.g., over 10% of the frames of a 6 second clips, over several hundred milliseconds, or over some other time duration). If a threshold percentage of the frames (e.g., 30-100%) exhibit a boost above the specified threshold, then the rolling shutter detection module 324 concludes that a rolling shutter effect is present for a set of frames. If the threshold percentage of frames is not met, the rolling shutter detection module 324 concludes that a rolling shutter effect is not present for the set of frames.
  • the rolling shutter detection module 324 detects rolling shutter effects using a number of different homographic mixture models Ft 2 ' ⁇ where each of the homographic mixture models varies with respect to the regularizer ⁇ .
  • a sufficiently high regularizer causes the homographic mixture model to be rigid and thus identical to the homographic model.
  • Relatively low regularizer values e.g., 3x10 ⁇ 5
  • Relatively high value regularizer values e.g., 4.7xl0 "4
  • model relatively slower moving distortions e.g., a person walking with a camera, a video shot from a boat).
  • the motion estimation module 310 estimates a number (e.g., four) homographic mixture models for the frame pair, where each homographic mixture model has a different
  • the rolling shutter detection module 324 determines the grid coverage Of 2 ' ⁇ and boost rse t, for each of the homographic mixture models with respect to the estimated homographic model Due to the difference in the regularizer, each homographic mixture model F/ 2 ' L) will have a different boost rse t ,x.
  • the homographic mixture models that meet the various boost thresholds are compared. In one embodiment, if a percentage (e.g., 5-15%, or higher) of the frames of the set meet one of the boost thresholds, then it is determined that a rolling shutter effect is present.
  • the rolling shutter correction module 324 is configured to alter the set of valid estimated motion models received from motion estimation module 310.
  • the valid motion models for that set of frames is permitted into include previously determined valid estimated homographic mixture models 3) for those frames. If a rolling shutter effect is determined not to be present across a set of frames, the valid estimated motion models for that set of frames is constrained to motion models having eight DOF (e.g., homographic models ) ), or lower.
  • the valid motion models for that set of frames are upgraded such that homographic mixture models 3) are considered valid for all frames in the set.
  • the rolling shutter correction module 324 determines first whether or not a rolling shutter effect is present, and if a rolling shutter effect is present, which of the homographic mixture models to use for a set of frames. As above, if a rolling shutter effect is determined not to be present, across a set of frames, the valid estimated motion models for that set of frames is constrained to motion models having eight DOF (e.g., homographic models ) ), or lower. If a rolling shutter effect is determined to be present, the homographic mixture model F 2 ' ⁇ used is the homographic mixture model meeting the boost threshold for the specified percentage of frames in the set.
  • the rolling shutter correction module 324 uses the homographic mixture model with the weakest regularizer for the frames in the set. As above, depending upon the implementation this homographic mixture model may be used for all frames in the set or only for those frames where the estimated homographic mixture model for that frame was determined to be valid.
  • the overlay and blur correction module 326 flags frames (or frame pairs) exhibiting a large amount of blur or significant static overlay, and is one means for performing this function.
  • the flags are used to place restrictions on the camera path itself and/or its use in generating the stabilized video.
  • a static overlay is identified in a frame by identifying those tracked features T ti exhibiting near zero motion M t (e.g., less than 0.2 pixels) as well as significantly small relative motion with respect to the dominant camera translation (e.g., ⁇ 20%). These tracked features are indicated to be static.
  • the overlay and blur correction module 326 aggregates the determinations that individual tracked features are static to determine whether a frame as a whole has a static overlay. To do this, the frame is divided into cells using a grid as described above. If more than 30% of a cell's tracked features are indicated as static, the cell is determined to be static.
  • a cell at given time t is indicated as being an overlay, that indication is propagated to a number of nearby frames (e.g., 30), as static overlays are typically present for more than a fraction of a second. If a sufficient number of cells of the grid are indicated as having an overlay, the entire frame is flagged as containing an overlay. This process is repeated for the other nearby frames, which may be similarly flagged. These flags indicate the presence of a static overlay, which may be taken into account in generating the stabilized video, described further below.
  • a number of nearby frames e.g. 30
  • the camera path stabilization module 330 generates a smoothed camera path and a crop transform (or simply crop), and is one means for performing this function.
  • the camera path stabilization module 330 receives as input the tracked features T and motions M generated by motion estimation 310, the set of valid estimated motion models F as generated by the motion estimation module 310 and as refined by the camera path analysis module 320, and any flags generated by the camera path analysis module 320.
  • the camera path 306 may be output separately.
  • This output camera path 306 may include the estimated motion models and/or the smoothed path and crop generated by the camera path stabilization module 330.
  • the smoothed camera path and crop can also used as an input to the stabilized video module 340 to generate a stabilized video 304.
  • the camera path stabilization module 330 includes a camera path smoothing module 332 and a cropping module 334.
  • the camera path smoothing module 332 smoothes the camera path by generating a smoothed path P that eliminates shake due to similarity (4 DOF), and lower DOF camera motions.
  • the smooth path P does not take into account or correct higher DOF (e.g. more than 4) motion.
  • the camera path smoothing module 332 generates the smoothed path of a frame at time P t using an LI path stabilization and the estimated valid translation similarity and identity motion models, and is one means for performing this function.
  • the camera path P t at time t is calculated using
  • P t includes a series of segments, each segment being one of a constant, linear, and/or parabolic motion. To accomplish this segmentation, P t is estimated by using a constrained LI optimization
  • B t represents the crop transform to be applied to the frame at time t to make a stabilized video 304 generated using the camera path appear if it was captured along the smooth path P t , thereby stabilizing the video.
  • the CLP Computer Infrastructure for Operations Research (COIN-OR) Linear Programming) simplex solver is used to determine B t . Crop determination is further described with respect to the cropping module 334, below.
  • camera path smoothing module 332 is configured to output the smoothed camera path P t based on the crop B t from the cropping module 334 and based on the estimated similarity motion model from each frame pair. If the similarity model for a given frame is not valid, a translation or identity model can be used in place of the similarity model in determining the smoothed path and crop.
  • the cropping module 334 is configured to determine the crop B t of each frame, and is one means for performing this function.
  • the crop governs the size of the frame.
  • the crop B t is determined by the camera motions present in the video.
  • the cropping module 334 is configured to find a crop B t that crops the content of each frame such that the remaining portion of the frame has the freedom to compensate for unwanted motion by adjusting what part of each frame is shown. Although larger crops generally make this easier, very large crops have the effect of removing frame content without providing additional stabilization benefit.
  • the cropping module determines the crop B t using Eq. (14) and by testing out several different crop window sizes to determine the crop B t that at least approximately minimizes Oj(P t ) of Eq. (14), where i represents the z ' -th crop test.
  • the crops tested include a 95% crop , a 90% crop, an 85% crop, and so on, down to a lower threshold such as a 70% crop.
  • the optimal crop size c opt is a percentage of the frame rectangle.
  • the crop transform B t is independent of crop size.
  • the determination of the optimal crop c opt for a temporally subsampled frame is, rather than being based on Eq. (14), is instead based on:
  • the determination of crop transform B t includes a number of constraints. First, the four corners of a crop window have a predetermined size less than the frame size. The size of the corners is determined to remain within in the frame after the transformation, e.g.,
  • a inequality constraint is placed, such that P t preserves a portion (e.g., 60%) of the original camera motion, thereby suppressing the perceived blur in the result at the cost of more shakiness. This may be isolated to one frame, or spread across several adjacent frames.
  • the scale of the c opt of the crop is added, with a small negative weight, to the objective as described in Eq. (14), effectively applying an inverse spring force on the crop window to bias the result towards less cropping.
  • the crop transform B t is determined on clips (or portions) of the video at a time (e.g., 6 seconds at a time). Additional constraints may be placed on individual clips that are not necessarily applicable to future clips.
  • the crop window is biased to be axis aligned and frame centered for the first frame of a clip (e.g., zero translation, a scale of 1, and zero rotation). This constrains the initial orientation for the crop of a clip.
  • the identity model is embedded in the similarity model and the crop transform is centered for the first frame of the clip.
  • the identity model is embedded in the similarity and change in rotation and scale of the crop across frames of that clip is set to zero (e.g., only translational DOFs are allowed).
  • the stabilization video module 340 is configured to generate a stabilized video 304 using the set of valid motion models F ⁇ and crop transform B t from each frame pair, and is one means for performing this function.
  • the stabilized video module 340 to generate the stabilized video 304, the stabilized video module 340 generates a stabilized frame J t for each input frame I t from the original input video 302.
  • the stabilized video module 340 generates each stabilized video frame J t by resampling the original frames I t according to the crop B t , and by correcting the resampling to account for any residual motion according to:
  • this resampling does not take into account higher DOF camera motions, such as those captured by the homographic and homographic mixture models. If no further correction were performed, such higher DOF motions would appear as high frequency residual wobble distortions in the resulting stabilized frames.
  • the additional terms H t and R t account for such higher DOF motions on a frame by frame basis. They affect the output frames J t where there is a homographic and/or homographic mixture model that has been determined to be valid for that frame.
  • Eq. (17) is recursively expanded as:
  • J t (x) I t (B t x) such that t £ ⁇ p,n ⁇ .
  • Eq. (19) is used to recursively compute the resampling location y t (p) from p to t and j ⁇ using a backward chain from n to t.
  • the two resampling locations are then linearly blended (or interpolated) to determine the final value of Jt(x), such that
  • the stabilized video module 340 generates the frames J t by applying the crop B t and the estimated valid motion models F t directly to the pixels of each frame I t .
  • the estimated motion models dictate a location where each pixel from each frame will appear after stabilization, if at all, as dictated by the crop. This process may be
  • the shake detection module 360 is configured to analyze videos to determine whether or not a video 302 would benefit from stabilization, as not all videos will benefit.
  • the process of determining whether or not a video would benefit from stabilization is referred to as camera shake detection, or simply shake detection.
  • the shake detection module 360 is configured to quantify an amount of shake in a video by generating a number of shake features.
  • the shake features are used to determine whether or not to stabilize the video 302.
  • Shake detection may be performed automatically or subject to a received request. Responsive to performing shake detection, a conclusion may be reached regarding whether the video has enough shake relative to a threshold to merit stabilization. Stabilization may be performed automatically upon reaching the threshold, or alternatively the user inputting the video 302 may be prompted with the option of performing stabilization based on the
  • the threshold for determining that a video would benefit from stabilization may vary between implementations. For videos with very little camera motion (or shake),
  • stabilization may actually make the video worse (e.g., more difficult to watch as a viewer) than if no stabilization were performed.
  • the threshold may be set such that stabilization is only performed if it improves the video. Processing costs involved with performing
  • stabilization may also be a factor.
  • the threshold may also be set such that stabilization is only performed if it improves the video enough to justify the processing cost. Thus, the threshold for determining whether to apply stabilization may vary between implementations.
  • the shake detection module 360 is configured to quantify the shake present in the video by generating a number of shake features. To generate the shake features, the shake detection module 360 generates a number of spectrograms S for the video 302 based on the estimated similarity models C t for the frames of the video (see Eq. (12)).
  • Each spectrogram S describes the frequency (or energy) components of the value of a single DOF of the similarity model across a number of adjacent frames.
  • each spectrogram represents either a DOF of translation t x along the x coordinate axis, a DOF of translation t y along the y coordinate axis, a DOF of scale change, or a DOF of rotation change.
  • the value of each DOF for a frame is represented by a parameter in the motion model, thus the value of each similarity DOF is the value of the corresponding parameter in the estimated similarity motion model FP for that frame.
  • Each spectrogram S also covers a limited time window of frames (e.g., 128 frames, or about 5 seconds of video.
  • the spectrograms also partially overlap with each other in time, such that two spectrograms may share frames.
  • a first spectrogram may be based on frames 0-128, a second spectrogram may be based on frames 64-196, and a third spectrogram may be based on frames 128-256.
  • the spectrogram S is generated in a frequency coordinate system where the portion of the spectrogram S k for each frame k is generated using the DOF values across the frames of the window and using a Fourier transform such as the Discrete Cosine Transform (DCT)-II algorith
  • DCT Discrete Cosine Transform
  • d represents the amount of contribution of a particular frequency/energy to the DOF values for the frames of the window.
  • An individual portion of the spectrogram S k can be stored, in the data storage 350, as a histogram comprising 128 bins, each bin representing a particular frequency/energy range. Each bin has a height of d n representing that bin's contribution to the DOF values of the frames of the window. Thus, in S k a comparatively tall bin indicates that the
  • Spectrogram S aggregates the DOF values for the frames of a time window into a histogram having a number of bins, where each bin represents a different frequency (or energy) range's contribution to the DOF's value for the frames in the window. [0105] Spectrograms may be compressed to help save memory space.
  • a scale 2 compression is used as it is generally expected that most energy in most video 302 spectrograms will be found at lower energies.
  • the contributions d n of similar energy ranges are aggregated together.
  • each portion of the spectrogram S k instead has 8 d n values, one for each energy bin.
  • FIG. 9 is an illustration of a number of spectrograms for a number of time windows and for the different degrees of freedom of the similarity model, according to one embodiment.
  • FIG. 9A illustrates the spectrograms of a first, short-length video 12 windows in length
  • FIG. 9B illustrates the spectrograms of a second, longer-length video 40 windows in length.
  • a separate graph is illustrated for each DOF of the similarity model for each video.
  • the example graphs of FIG. 9 assume 128 frames per spectrogram, scale 2 compression and thus 8 energy bins per spectrogram, and approximately 50% window overlap in the frames of each spectrogram.
  • the y axis of each graph illustrates the 8 energy bins, with bin number increasing with respect to energy.
  • the x axis of each graph illustrates the spectrograms of the video by window.
  • the color of each pixel of the graph represents the amount of energy (i.e., motion) within a particular frequency range within each window of frames
  • the shake features may be generated from the spectrograms using any one of several different methods including, for example, based on the mean, median, and/or maximum of spectrogram histogram bin height across all windows and based on a separate histogram that groups the spectrogram's energy according to percentile.
  • a separate histogram that groups the spectrogram's energy according to percentile.
  • One or more sets of shake features may be generated from the spectrograms by taking one or more of the mean, maximum, and median spectrogram height of each bin across the windows of a video.
  • the height of a bin of a spectrogram represents the contribution of particular range of energies/frequencies to the DOF values of the windows on a window per window basis.
  • the mean across all windows represents the average contribution of that bin's frequencies/energies to the video as a whole by window across the windows of the video.
  • the maximum across all windows represents the maximum contribution of that bin's frequencies/energies to the video as a whole by window across the windows of the video
  • the median across all windows represents the median contribution of that bin's frequencies/energies to the video as a whole by window across the windows of the video.
  • Another set of shake features may be generated by from the spectrograms by creating a separate set of histograms of the spectrogram domain, one domain histogram for each energy bin for each DOF, and thus using the exemplary conditions above, 32 domain histograms total (e.g., 8 energy bin times 4 DOF).
  • Each domain histogram has a number of bins (referred to as domain bins to avoid confusion with the energy bins of the underlying spectrograms).
  • Each domain bin has its own shake feature. Continuing with the example from above, if each domain histogram has 10 domain bins, then the shake features generated by this technique number 320 in total.
  • a domain histogram groups the heights/contributions, dge, of the individual windows of a single energy bin (e.g., one of 0-7) of spectrogram into percentile ranges of contribution relative to all energy bins of the spectrogram across all windows.
  • the domain histogram is normalized on a scale of , for example, [0,1], where 0 represents a contribution value d n of zero, or alternatively the lowest amount of contribution d n , min in the spectrogram, and 1 represents the highest amount of contribution d n , max in the spectrogram.
  • Each domain bin covers a defined percentile range of contribution values.
  • the height of each domain bin is the number of windows in the energy bin having that contribution values d n within that percentile range. For example, if each of 10 domain bins covers a 10% range, a height of a first domain bin indicates the number of windows of the energy bin (e.g., spectrogram bin 0 out of bins 0-7) having contribution values d n between 0- 10% the contribution of the maximum contribution of any bin in the spectrogram.
  • a height of a second domain bin indicates the number of windows of that same energy bin (e.g., again spectrogram bin 0) having contribution values d n between 11-20% the contribution of the maximum contribution of any bin in the spectrogram.
  • the heights of the domain bins may be normalized by the total number of windows in the video so that the domain bins are invariant with respect to the length of the video. This allows domain bin shake features from videos of various lengths to be compared despite having differing numbers of windows.
  • the shake features are analyzed to determine whether to apply stabilization to a video.
  • the shake detection module 360 uses machine learning algorithms to trains a shake classifier to determine whether to apply stabilization.
  • the shake detection module 360 uses shake features from known videos and determinations (e.g., yes, no) of whether these known videos would be stabilized as training inputs.. By training the classifier with decisions about whether these known videos would or would not be stabilized, the shake classifier is trained to learn whether or not later received videos 302 should be stabilized.
  • the shake features used to train the shake classifier may vary between
  • 32 mean shake features, 32 maximum shake features, and 320 domain shake features are used to train the classifier.
  • any combination of mean, max, median, and domain shake features may be used to train the classifier.
  • additional features of the video may also be used to train the classifier. These features may include, for example, features deduced from the blur present in the video, as well as non-shake features such as the scene content of the video, and the audio of the video.
  • a video 302 may be analyzed to determine whether or not to stabilize the video.
  • the shake detection module 360 processes the video to generate shake features as described above.
  • the shake features (and any other features are input to the shake classifier.
  • the shake classifier then outputs a determination of whether or not the video should be stabilized. Stabilization may then be automatically conducted, or conducted responsive to a user input.
  • FIG. 4 is a flowchart illustrating a process for determining a camera path of a video, according to one embodiment.
  • the stabilization server 112 accesses 402 a video and generates 404 two dimensional tracked features for at least two adjacent frames of the received video.
  • the tracked features of the adjacent frames indicate an inter- frame motion of the camera.
  • a number of different motion models are each individually applied 406 to the tracked features of a frame to determine properties of the motion models. Each motion model has a different number of DOF. Based on the properties, a determination 408 is made regarding which of the motion models are valid.
  • a camera path 410 describing the motion of the camera used to capture the video is generated based on the motion models that are valid for the inter-frame motion between the adjacent frames.
  • FIG. 5 is a flowchart illustrating a process for detecting and correcting rolling shutter in a video, according to one embodiment.
  • the stabilization server access 502 a video and generates 504 two dimensional tracked features for at least two adjacent frames of the received video.
  • the tracked features of the adjacent frames indicate an inter- frame motion of the camera.
  • a homographic model is applied 506 to the inter-frame motion to determine a number of tracked features that are inliers matching the homographic model.
  • a homographic mixture model is applied 508 to the inter-frame motion to determine a number of tracked features that are inliers matching the homographic mixture model.
  • a stabilized is generated by applying the homographic mixture model to the adjacent frames of the video.
  • FIG. 8 is a flowchart illustrating a process for detecting camera shake in a video, according to one embodiment.
  • the stabilization system 112 accesses 802 a video and estimates 804, for a number of frames of the video, values (or parameters) of the DOFs of a similarity motion model as described above.
  • the stabilization system 112 generates 806 a spectrogram for each DOFs and time window, such that each spectrogram is based on the values of the DOFs over a time window comprising a plurality of adjacent frames of the video.
  • the stabilization system 112 uses the spectrograms, the stabilization system 112 generates 808 shake features based on the spectrograms.
  • the stabilization system 112 classifies 810 the video based on the shake features and a previously trained shake classifier.
  • the shake classifier classifies 810 the video into one of two categories, videos that should be stabilized and videos that should not be stabilized.
  • the stabilization system 812 stabilizes the video based on the classification, either automatically or responsive to a user input.

Landscapes

  • Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Signal Processing (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Studio Devices (AREA)
  • Image Analysis (AREA)

Abstract

La présente invention se rapporte à un système de stabilisation de vidéo en ligne, simple à utiliser. L'invention se rapporte d'autre part à des procédés pour l'utilisation de ce système. Des vidéos sont stabilisées après capture, et la stabilisation peut donc fonctionner sur n'importe quelle forme de vidéo, qu'elle soit patrimoniale ou récemment capturée. Dans l'un des modes de réalisation de la présente invention, le système de stabilisation de vidéo est entièrement automatique. Il ne nécessite aucune entrée ni aucun réglage de paramètres de la part de l'utilisateur autre que la vidéo en tant que telle. Le système de stabilisation de vidéo utilise un modèle de mouvements en cascade afin de choisir la correction qui doit être appliquée à différentes trames d'une vidéo. Dans divers modes de réalisation, le système de stabilisation de vidéo est apte à détecter et à corriger des artefacts de gigue haute fréquence, des artefacts de tremblement basse fréquence, des artefacts de roulement d'obturateur, des mouvements significatifs en arrière-plan, un éclairage de mauvaise qualité, des coupes dans une scène, et des vidéos de longue dimension et de courte dimension.
PCT/US2014/023211 2013-03-15 2014-03-11 Estimation de mouvements en cascade de caméra, détection de roulement d'obturateur, et détection de tremblement de caméra pour améliorer la stabilisation d'une vidéo WO2014150421A1 (fr)

Priority Applications (5)

Application Number Priority Date Filing Date Title
EP20209572.5A EP3800878B1 (fr) 2013-03-15 2014-03-11 Estimation de mouvements en cascade de caméra, détection de roulement d'obturateur, et détection de tremblement de caméra pour améliorer la stabilisation d'une vidéo
EP14768016.9A EP2974274B1 (fr) 2013-03-15 2014-03-11 Estimation de mouvements en cascade de caméra, détection de roulement d'obturateur, et détection de tremblement de caméra pour améliorer la stabilisation d'une vidéo
KR1020157029569A KR102185963B1 (ko) 2013-03-15 2014-03-11 비디오 안정화를 위한 캐스케이드 카메라 모션 추정, 롤링 셔터 검출 및 카메라 흔들림 검출
CN201480016067.2A CN105052129B (zh) 2013-03-15 2014-03-11 级联相机运动估计、滚动快门检测和用于视频稳定性的相机抖动检测
KR1020207034211A KR102225410B1 (ko) 2013-03-15 2014-03-11 비디오 안정화를 위한 캐스케이드 카메라 모션 추정, 롤링 셔터 검출 및 카메라 흔들림 검출

Applications Claiming Priority (4)

Application Number Priority Date Filing Date Title
US201361799985P 2013-03-15 2013-03-15
US61/799,985 2013-03-15
US13/854,819 US9374532B2 (en) 2013-03-15 2013-04-01 Cascaded camera motion estimation, rolling shutter detection, and camera shake detection for video stabilization
US13/854,819 2013-04-01

Publications (1)

Publication Number Publication Date
WO2014150421A1 true WO2014150421A1 (fr) 2014-09-25

Family

ID=51525713

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2014/023211 WO2014150421A1 (fr) 2013-03-15 2014-03-11 Estimation de mouvements en cascade de caméra, détection de roulement d'obturateur, et détection de tremblement de caméra pour améliorer la stabilisation d'une vidéo

Country Status (5)

Country Link
US (3) US9374532B2 (fr)
EP (2) EP3800878B1 (fr)
KR (2) KR102225410B1 (fr)
CN (1) CN105052129B (fr)
WO (1) WO2014150421A1 (fr)

Families Citing this family (49)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8896715B2 (en) * 2010-02-11 2014-11-25 Microsoft Corporation Generic platform video image stabilization
US9824426B2 (en) 2011-08-01 2017-11-21 Microsoft Technology Licensing, Llc Reduced latency video stabilization
US9374532B2 (en) 2013-03-15 2016-06-21 Google Inc. Cascaded camera motion estimation, rolling shutter detection, and camera shake detection for video stabilization
US9336460B2 (en) * 2013-05-31 2016-05-10 Intel Corporation Adaptive motion instability detection in video
US9277129B2 (en) * 2013-06-07 2016-03-01 Apple Inc. Robust image feature based video stabilization and smoothing
US9210327B2 (en) * 2013-12-02 2015-12-08 Yahoo! Inc. Blur aware photo feedback
US9939253B2 (en) * 2014-05-22 2018-04-10 Brain Corporation Apparatus and methods for distance estimation using multiple image sensors
CN105493496B (zh) 2014-12-14 2019-01-18 深圳市大疆创新科技有限公司 一种视频处理方法、装置及图像系统
US9418396B2 (en) 2015-01-15 2016-08-16 Gopro, Inc. Watermarking digital images to increase bit depth
US9877036B2 (en) 2015-01-15 2018-01-23 Gopro, Inc. Inter frame watermark in a digital video
GB2523253B (en) * 2015-01-23 2017-04-12 Visidon Oy Image processing method
US9525821B2 (en) 2015-03-09 2016-12-20 Microsoft Technology Licensing, Llc Video stabilization
US10708571B2 (en) 2015-06-29 2020-07-07 Microsoft Technology Licensing, Llc Video frame processing
US20170006219A1 (en) 2015-06-30 2017-01-05 Gopro, Inc. Image stitching in a multi-camera array
CN105163046B (zh) * 2015-08-17 2018-11-06 成都泛视微星科技有限公司 一种基于网格吸点非参数化运动模型的视频防抖方法
US10602153B2 (en) 2015-09-11 2020-03-24 Facebook, Inc. Ultra-high video compression
US10375156B2 (en) 2015-09-11 2019-08-06 Facebook, Inc. Using worker nodes in a distributed video encoding system
US10499070B2 (en) 2015-09-11 2019-12-03 Facebook, Inc. Key frame placement for distributed video encoding
US10602157B2 (en) 2015-09-11 2020-03-24 Facebook, Inc. Variable bitrate control for distributed video encoding
US10341561B2 (en) * 2015-09-11 2019-07-02 Facebook, Inc. Distributed image stabilization
US10506235B2 (en) 2015-09-11 2019-12-10 Facebook, Inc. Distributed control of video encoding speeds
US10063872B2 (en) 2015-09-11 2018-08-28 Facebook, Inc. Segment based encoding of video
US10044944B2 (en) 2015-09-28 2018-08-07 Gopro, Inc. Automatic composition of video with dynamic background and composite frames selected based on foreground object criteria
GB2549074B (en) 2016-03-24 2019-07-17 Imagination Tech Ltd Learned feature motion detection
US10045120B2 (en) 2016-06-20 2018-08-07 Gopro, Inc. Associating audio with three-dimensional objects in videos
US9749738B1 (en) 2016-06-20 2017-08-29 Gopro, Inc. Synthesizing audio corresponding to a virtual microphone location
US9922398B1 (en) 2016-06-30 2018-03-20 Gopro, Inc. Systems and methods for generating stabilized visual content using spherical visual content
US10134114B2 (en) 2016-09-20 2018-11-20 Gopro, Inc. Apparatus and methods for video image post-processing for segmentation-based interpolation
US10313686B2 (en) 2016-09-20 2019-06-04 Gopro, Inc. Apparatus and methods for compressing video content using adaptive projection selection
US10003768B2 (en) 2016-09-28 2018-06-19 Gopro, Inc. Apparatus and methods for frame interpolation based on spatial considerations
US10591731B2 (en) * 2016-12-06 2020-03-17 Google Llc Ocular video stabilization
US10489897B2 (en) 2017-05-01 2019-11-26 Gopro, Inc. Apparatus and methods for artifact detection and removal using frame interpolation techniques
KR102330264B1 (ko) * 2017-08-04 2021-11-23 삼성전자주식회사 움직임 정보에 기반하여 동영상을 재생하기 위한 장치 및 그의 동작 방법
AU2017245322A1 (en) * 2017-10-10 2019-05-02 Canon Kabushiki Kaisha Method, system and apparatus for selecting frames of a video sequence
US10740620B2 (en) 2017-10-12 2020-08-11 Google Llc Generating a video segment of an action from a video
US10740431B2 (en) 2017-11-13 2020-08-11 Samsung Electronics Co., Ltd Apparatus and method of five dimensional (5D) video stabilization with camera and gyroscope fusion
US10587807B2 (en) 2018-05-18 2020-03-10 Gopro, Inc. Systems and methods for stabilizing videos
US10750092B2 (en) 2018-09-19 2020-08-18 Gopro, Inc. Systems and methods for stabilizing videos
KR102573302B1 (ko) 2018-10-10 2023-08-31 삼성전자 주식회사 영상의 안정화를 위한 카메라 모듈, 그것을 포함하는 전자 장치 및 전자 장치의 영상 안정화 방법
CN109348125B (zh) * 2018-10-31 2020-02-04 Oppo广东移动通信有限公司 视频校正方法、装置、电子设备和计算机可读存储介质
US10482584B1 (en) * 2019-01-31 2019-11-19 StradVision, Inc. Learning method and learning device for removing jittering on video acquired through shaking camera by using a plurality of neural networks for fault tolerance and fluctuation robustness in extreme situations, and testing method and testing device using the same
US11089220B2 (en) 2019-05-02 2021-08-10 Samsung Electronics Co., Ltd. Electronic test device, method and computer-readable medium
EP3987763A4 (fr) * 2019-06-21 2023-07-12 GoPro, Inc. Systèmes et procédés pour stabiliser des vidéos
KR102176273B1 (ko) * 2019-07-04 2020-11-09 재단법인대구경북과학기술원 동영상 수평 조정 시스템, 방법 및 컴퓨터 프로그램
US11599974B2 (en) * 2019-11-22 2023-03-07 Nec Corporation Joint rolling shutter correction and image deblurring
US11694311B2 (en) * 2020-03-04 2023-07-04 Nec Corporation Joint rolling shutter image stitching and rectification
CN111709979B (zh) * 2020-05-15 2023-08-25 北京百度网讯科技有限公司 图像对齐的方法、装置、电子设备和存储介质
CN114095659B (zh) * 2021-11-29 2024-01-23 厦门美图之家科技有限公司 一种视频防抖方法、装置、设备及存储介质
CN115439501B (zh) * 2022-11-09 2023-04-07 慧视云创(北京)科技有限公司 一种视频流动态背景构造方法、装置及检测运动目标方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7433497B2 (en) * 2004-01-23 2008-10-07 Hewlett-Packard Development Company, L.P. Stabilizing a sequence of image frames
US20090066800A1 (en) * 2007-09-06 2009-03-12 Texas Instruments Incorporated Method and apparatus for image or video stabilization
JP2010252325A (ja) * 2009-04-16 2010-11-04 Nvidia Corp 画像補正のためのシステム及び方法
US20120105654A1 (en) * 2010-10-28 2012-05-03 Google Inc. Methods and Systems for Processing a Video for Stabilization and Retargeting

Family Cites Families (75)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5903454A (en) * 1991-12-23 1999-05-11 Hoffberg; Linda Irene Human-factored interface corporating adaptive pattern recognition based controller apparatus
JP3626218B2 (ja) * 1993-08-23 2005-03-02 ソニー株式会社 動き量検出装置及び動き量検出方法
AU5273496A (en) * 1995-03-22 1996-10-08 Idt International Digital Technologies Deutschland Gmbh Method and apparatus for coordination of motion determination over multiple frames
WO2001069932A1 (fr) * 2000-03-10 2001-09-20 Sensormatic Electronics Corporation Procede et appareil pour la poursuite et la detection d'un objet
US6965645B2 (en) * 2001-09-25 2005-11-15 Microsoft Corporation Content-based characterization of video frame sequences
KR101036787B1 (ko) * 2005-01-14 2011-05-25 가부시키가이샤 모르포 움직임 벡터 연산 방법과 이 방법을 이용한 손 떨림 보정장치, 촬상 장치, 및 동영상 생성 장치
US20060215036A1 (en) * 2005-03-25 2006-09-28 Multivision Intelligent Surveillance (Hk) Ltd. Method and apparatus for video stabilization
US7558405B2 (en) * 2005-06-30 2009-07-07 Nokia Corporation Motion filtering for video stabilization
US8666661B2 (en) * 2006-03-31 2014-03-04 The Boeing Company Video navigation
JP4961850B2 (ja) * 2006-06-15 2012-06-27 ソニー株式会社 動き検出方法、動き検出方法のプログラム、動き検出方法のプログラムを記録した記録媒体及び動き検出装置
US8073196B2 (en) * 2006-10-16 2011-12-06 University Of Southern California Detection and tracking of moving objects from a moving platform in presence of strong parallax
US8019180B2 (en) * 2006-10-31 2011-09-13 Hewlett-Packard Development Company, L.P. Constructing arbitrary-plane and multi-arbitrary-plane mosaic composite images from a multi-imager
WO2008108071A1 (fr) * 2007-03-06 2008-09-12 Panasonic Corporation Appareil et procédé de traitement d'image, programme de traitement d'image et processeur d'image
KR100866963B1 (ko) * 2007-03-12 2008-11-05 삼성전자주식회사 수평 방향의 기울어짐 왜곡과 수직 방향의 스케일링 왜곡을보정할 수 있는 디지털 영상 안정화 방법
JP4958610B2 (ja) * 2007-04-06 2012-06-20 キヤノン株式会社 画像防振装置、撮像装置及び画像防振方法
EP2151128A4 (fr) * 2007-04-25 2011-11-16 Miovision Technologies Inc Procédé et système pour analyser un contenu multimédia
KR101392732B1 (ko) * 2007-08-20 2014-05-08 삼성전자주식회사 손떨림에 의한 움직임 추정 장치 및 방법, 그를 이용한영상 촬상 장치
JP2009077362A (ja) * 2007-08-24 2009-04-09 Sony Corp 画像処理装置、動画再生装置、これらにおける処理方法およびプログラム
TWI381719B (zh) * 2008-02-18 2013-01-01 Univ Nat Taiwan 穩定全幅式視訊之方法
US8494058B2 (en) * 2008-06-23 2013-07-23 Mediatek Inc. Video/image processing apparatus with motion estimation sharing, and related method and machine readable medium
EP2157800B1 (fr) * 2008-08-21 2012-04-18 Vestel Elektronik Sanayi ve Ticaret A.S. Procédé et appareil pour augmenter la fréquence de trames d'un signal vidéo
US8102428B2 (en) * 2008-08-28 2012-01-24 Adobe Systems Incorporated Content-aware video stabilization
GB0818561D0 (en) * 2008-10-09 2008-11-19 Isis Innovation Visual tracking of objects in images, and segmentation of images
JP2010097056A (ja) * 2008-10-17 2010-04-30 Seiko Epson Corp 表示装置
JP5284048B2 (ja) * 2008-11-12 2013-09-11 キヤノン株式会社 画像処理装置、撮像装置及び画像処理方法
US8411966B2 (en) * 2009-03-10 2013-04-02 Her Majesty The Queen In Right Of Canada, As Represented By The Minister Of Industry, Through The Communications Research Centre Canada Estimation of image relations from point correspondences between images
KR101614914B1 (ko) * 2009-07-23 2016-04-25 삼성전자주식회사 모션 적응적 고대비 영상 획득 장치 및 방법
US9626769B2 (en) * 2009-09-04 2017-04-18 Stmicroelectronics International N.V. Digital video encoder system, method, and non-transitory computer-readable medium for tracking object regions
WO2011036625A2 (fr) * 2009-09-23 2011-03-31 Ramot At Tel-Aviv University Ltd. Système, procédé et produit de programme d'ordinateur pour la détection de mouvement
US8135221B2 (en) * 2009-10-07 2012-03-13 Eastman Kodak Company Video concept classification using audio-visual atoms
US9667887B2 (en) * 2009-11-21 2017-05-30 Disney Enterprises, Inc. Lens distortion method for broadcast video
US8553982B2 (en) * 2009-12-23 2013-10-08 Intel Corporation Model-based play field registration
US8179446B2 (en) * 2010-01-18 2012-05-15 Texas Instruments Incorporated Video stabilization and reduction of rolling shutter distortion
US8358359B2 (en) * 2010-01-21 2013-01-22 Microsoft Corporation Reducing motion-related artifacts in rolling shutter video information
US8896715B2 (en) * 2010-02-11 2014-11-25 Microsoft Corporation Generic platform video image stabilization
US8350922B2 (en) * 2010-04-30 2013-01-08 Ecole Polytechnique Federale De Lausanne Method to compensate the effect of the rolling shutter effect
US20120010513A1 (en) * 2010-07-08 2012-01-12 Wong Stephen T C Chemically-selective, label free, microendoscopic system based on coherent anti-stokes raman scattering and microelectromechanical fiber optic probe
US8571328B2 (en) * 2010-08-16 2013-10-29 Adobe Systems Incorporated Determining correspondence between image regions
US8872928B2 (en) * 2010-09-14 2014-10-28 Adobe Systems Incorporated Methods and apparatus for subspace video stabilization
US8810692B2 (en) * 2010-10-19 2014-08-19 Apple Inc. Rolling shutter distortion correction
US9538982B2 (en) * 2010-12-18 2017-01-10 Massachusetts Institute Of Technology User interface for ultrasound scanning system
US20120182442A1 (en) * 2011-01-14 2012-07-19 Graham Kirsch Hardware generation of image descriptors
US8964041B2 (en) * 2011-04-07 2015-02-24 Fr Vision Ab System and method for video stabilization of rolling shutter cameras
US8724854B2 (en) * 2011-04-08 2014-05-13 Adobe Systems Incorporated Methods and apparatus for robust video stabilization
WO2013005316A1 (fr) * 2011-07-06 2013-01-10 株式会社モルフォ Dispositif de traitement d'image, procédé de traitement d'image et programme de traitement d'image
US8718378B2 (en) * 2011-07-11 2014-05-06 Futurewei Technologies, Inc. Image topological coding for visual search
US8913140B2 (en) * 2011-08-15 2014-12-16 Apple Inc. Rolling shutter reduction based on motion sensors
TWI478833B (zh) * 2011-08-31 2015-04-01 Autoequips Tech Co Ltd 調校車用影像裝置之方法及其系統
JP5729237B2 (ja) * 2011-09-26 2015-06-03 カシオ計算機株式会社 画像処理装置、画像処理方法、及びプログラム
US8699852B2 (en) * 2011-10-10 2014-04-15 Intellectual Ventures Fund 83 Llc Video concept classification using video similarity scores
US8903043B2 (en) * 2011-10-24 2014-12-02 Bruker Axs, Inc. Method for correcting timing skew in X-ray data read out of an X-ray detector in a rolling shutter mode
US8457357B2 (en) * 2011-11-09 2013-06-04 Disney Enterprises, Inc. Relative pose estimation of non-overlapping cameras using the motion of subjects in the camera fields of view
TWI469062B (zh) * 2011-11-11 2015-01-11 Ind Tech Res Inst 影像穩定方法及影像穩定裝置
US8842883B2 (en) * 2011-11-21 2014-09-23 Seiko Epson Corporation Global classifier with local adaption for objection detection
US9003289B2 (en) * 2012-02-23 2015-04-07 Google Inc. Automatic detection of suggested video edits
US20130251340A1 (en) * 2012-03-21 2013-09-26 Wei Jiang Video concept classification using temporally-correlated grouplets
US9232230B2 (en) * 2012-03-21 2016-01-05 Vixs Systems, Inc. Method and device to identify motion vector candidates using a scaled motion search
US9129524B2 (en) * 2012-03-29 2015-09-08 Xerox Corporation Method of determining parking lot occupancy from digital camera images
ITVI20120087A1 (it) * 2012-04-17 2013-10-18 St Microelectronics Srl Stabilizzazione video digitale
US8948497B2 (en) * 2012-09-04 2015-02-03 Digital Signal Corporation System and method for increasing resolution of images obtained from a three-dimensional measurement system
US8860825B2 (en) * 2012-09-12 2014-10-14 Google Inc. Methods and systems for removal of rolling shutter effects
US9684941B2 (en) * 2012-10-29 2017-06-20 Digimarc Corporation Determining pose for use with digital watermarking, fingerprinting and augmented reality
EP2739044B1 (fr) * 2012-11-29 2015-08-12 Alcatel Lucent Serveur de conférence vidéo avec détection de tremblement de caméra
TWI602152B (zh) * 2013-02-06 2017-10-11 聚晶半導體股份有限公司 影像擷取裝置及其影像處理方法
US9084531B2 (en) * 2013-02-27 2015-07-21 Siemens Aktiengesellschaft Providing real-time marker detection for a stent in medical imaging
US9317781B2 (en) * 2013-03-14 2016-04-19 Microsoft Technology Licensing, Llc Multiple cluster instance learning for image classification
US9374532B2 (en) * 2013-03-15 2016-06-21 Google Inc. Cascaded camera motion estimation, rolling shutter detection, and camera shake detection for video stabilization
US20140313325A1 (en) * 2013-04-18 2014-10-23 Ge Aviation Systems Llc Method of generating a spatial and spectral object model
JP6209002B2 (ja) * 2013-07-16 2017-10-04 キヤノン株式会社 撮像装置およびその制御方法
US9554048B2 (en) * 2013-09-26 2017-01-24 Apple Inc. In-stream rolling shutter compensation
US20160262685A1 (en) * 2013-11-12 2016-09-15 Highland Instruments, Inc. Motion analysis systemsand methods of use thereof
US9854168B2 (en) * 2014-03-07 2017-12-26 Futurewei Technologies, Inc. One-pass video stabilization
CN105095900B (zh) * 2014-05-04 2020-12-08 斑马智行网络(香港)有限公司 一种提取标准卡片中特定信息的方法和装置
FR3027144B1 (fr) * 2014-10-09 2016-11-04 St Microelectronics Sa Procede et dispositif de determination de mouvement entre des images video successives
US20170213576A1 (en) * 2016-01-22 2017-07-27 Artur Nugumanov Live Comics Capturing Camera

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7433497B2 (en) * 2004-01-23 2008-10-07 Hewlett-Packard Development Company, L.P. Stabilizing a sequence of image frames
US20090066800A1 (en) * 2007-09-06 2009-03-12 Texas Instruments Incorporated Method and apparatus for image or video stabilization
JP2010252325A (ja) * 2009-04-16 2010-11-04 Nvidia Corp 画像補正のためのシステム及び方法
US20120105654A1 (en) * 2010-10-28 2012-05-03 Google Inc. Methods and Systems for Processing a Video for Stabilization and Retargeting

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
MATTHIAS GRUNDMANN T AL.: "CALIBRATION-FREE RODOI:10.1109/ICCPHOT.2012.6215213", COMPUTATIONAL PHOTOGRAPHY (ICCP), 2012 IEEE INTERNATIONAL CONFERENCE., 28 April 2012 (2012-04-28) - 29 April 2012 (2012-04-29), pages 1 - 8, XP032185752, Retrieved from the Internet <URL:http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=6215213> DOI: 10.1109/ICCPHOT.2012.6215213 *
See also references of EP2974274A4 *

Also Published As

Publication number Publication date
KR20150132846A (ko) 2015-11-26
EP2974274A1 (fr) 2016-01-20
CN105052129A (zh) 2015-11-11
EP3800878A1 (fr) 2021-04-07
KR20200137039A (ko) 2020-12-08
KR102185963B1 (ko) 2020-12-03
US9888180B2 (en) 2018-02-06
US9374532B2 (en) 2016-06-21
EP2974274B1 (fr) 2020-12-09
EP2974274A4 (fr) 2016-11-09
US20140267801A1 (en) 2014-09-18
KR102225410B1 (ko) 2021-03-08
US20170195575A1 (en) 2017-07-06
EP3800878B1 (fr) 2024-01-31
CN105052129B (zh) 2019-04-23
US20160269642A1 (en) 2016-09-15
US9635261B2 (en) 2017-04-25

Similar Documents

Publication Publication Date Title
US9888180B2 (en) Cascaded camera motion estimation, rolling shutter detection, and camera shake detection for video stabilization
Liu et al. Meshflow: Minimum latency online video stabilization
Liu et al. Bundled camera paths for video stabilization
US9854168B2 (en) One-pass video stabilization
CN102742260B (zh) 通用平台视频图像稳定化
US20140072216A1 (en) Image De-Hazing by Solving Transmission Value
CN106210448B (zh) 一种视频图像抖动消除处理方法
Okade et al. Video stabilization using maximally stable extremal region features
Souza et al. Digital video stabilization based on adaptive camera trajectory smoothing
CN115606174A (zh) 视频质量评估方法和设备
Kim et al. Spatio-temporal weighting in local patches for direct estimation of camera motion in video stabilization
CN114429191B (zh) 基于深度学习的电子防抖方法、系统及存储介质
US20180103278A1 (en) Identification of captured videos
Chen et al. Pixstabnet: Fast multi-scale deep online video stabilization with pixel-based warping
CN112801890A (zh) 一种视频处理方法、装置及设备
Hara et al. A differentiable approximation approach to contrast-aware image fusion
e Souza et al. Improvement of global motion estimation in two‐dimensional digital video stabilisation methods
Peng et al. An efficient temporal distortion measure of videos based on spacetime texture
Unnikrishnan et al. Video stabilization performance enhancement for low-texture videos
Zhang et al. Photometric Stabilization for Fast‐forward Videos
Tian et al. Stabilized videos source identification method based on undecimated dual-tree complex wavelet transform and Bayesian adaptive direct search algorithm
US20240062338A1 (en) Systems and methods for blur identification and correction
Tian et al. Double-weighted photo response non-uniformity extraction model based on multi-scale neighborhood shrinkage and structural filtering for YouTube compressed videos
Zhou et al. A joint spatial-temporal 3D video stabilization algorithm
Shagrithaya et al. Video stabilization using sliding frame window

Legal Events

Date Code Title Description
WWE Wipo information: entry into national phase

Ref document number: 201480016067.2

Country of ref document: CN

121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 14768016

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 2014768016

Country of ref document: EP

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 20157029569

Country of ref document: KR

Kind code of ref document: A